Abstract
Current models of vision propose that the brain uses a multi-layered architecture to reduce the high dimensional input it receives (e.g. a street crowded with people and vehicles) to lower dimensional representations that support a variety of everyday categorization behaviors (e.g. perceiving "a happy Mary," on a "bus" in "New York City"). To explicitly study the reduction mechanisms from brain activity, on each experimental trial we used the Bubbles technique to randomly sample visual information from an ambiguous scene—i.e. Dali painting Slave Market with Disappearing Bust of Voltaire (see SuppFig A) — while simultaneously measuring the perceptual decisions (i.e. "the nuns," "Voltaire," "don't know") of 5 observers and their dynamic brain activity recorded with MEG. For each observer we first quantified the relationship between the random information samples and the perceptual decision across trials (2885-4145 per observer) with mutual information. This revealed the stimulus features that functionally link the painting to each perceptual decision—henceforth, the diagnostic features (see SuppFig A). A similar analysis performed on 12,000 MEG sources (between 0 and 400ms post-stimulus) revealed a consistent data reduction pattern in each observer: before 170ms, occipital cortex represents the diagnostic features of each perceptual decision but also other, nondiagnostic features; after 170ms, a reduction occurs at the juncture between occipital cortex and ventral regions whereby only diagnostic features remain represented in occipital and ventral cortex (SuppFig B). In ventral cortex, Diagnostic features reach the right fusiform gyrus around 140-170ms via a middle cluster of voxels associated with perceptual decisions. Diagnostic features then migrate to a second cluster, around 170-200ms, where we demonstrate, using information theoretic redundancy, that they form specific feature-based distributed representations that support perceptual decision behaviors (SuppFig C).
Meeting abstract presented at VSS 2018